Deep Learning Model Can Identify Hydroxychloroquine Retinal Toxicity

Ophtalmological practice, Geneva, Switzerland, Carrying out OCT angiography to detect the presence of neovascularisation, angiography with autofluorescence and optical coherence tomography. (Photo by: BSIP/Universal Images Group via Getty Images)
Investigators are creating software that can identify ellipsoid zone loss to screen patients for toxicity.

A deep learning algorithm can automatically, accurately estimate ellipsoid zone (EZ) loss to screen patients for toxicity associated with hydroxychloroquine (HCQ) use, according to research published in Ophthalmology Science.

HCQ-associated retinal toxicity  can eventually lead to central vision loss in approximately 7.5% of patients who take the drug for more than 10 years, according to researchers.

The study included data from a clinical study ( NCT01145196) of 85 patients (mean 59±12 years, 93% women) who used HCQ for 14±7.2 years. The researchers collected spectral domain optical coherence tomography (SD-OCT) images: training the mask-RCNN on individual 1 horizontal 30o scan and 1 volumetric macular cube OCT-B scans to construct an EZ loss map for each eye.

They compared the EZ loss area and the precision, recall and F1-score metrics with human grader annotations and the determination of the toxicity based on screening guidelines.

The combined model (CPN) had the best overall performance (precision 0.90±0.09, recall 0.88±0.08, F1 score = 0.89±0.07), performing better than the horizontal-only model (M-RCNNH) (precision 0.79±0.17, recall 0.96±0.04, IOU 0.78±0.15, F1 score 0.86±0.12) and vertical only (M-RCNNV) (precision 0.71±0.21, recall 0.94±0.06, IOU 0.69±0.21, F1 score 0.79±0.16) models. Its accuracy was not significantly different from human annotation (precision 0.85±0.09, recall 0.98±0.01, IOU 0.82±0.12, and F1 score 0.91±0.06).

Total EZ loss area is negatively correlated with Humphrey visual field mean deviation (R2=-0.81). The algorithm did not find substantial EZ loss regions in the unaffected eyes.

“Clinical translation of this tool would enable automated and objective identification of patients who demonstrate changes concerning for toxicity, which could aid the screening ophthalmologist,” according to the researchers. “Corroborating these results with ancillary functional testing and/or identifying patients who would benefit from referral to specialists would improve current screening methods.” 

The investigators believe implementing this algorithm could also help screening outside of ophthalmology offices, as OCTs become more ubiquitous in internal medicine settings, the report suggests. “Quantitative data produced from the algorithms could provide surrogate endpoints for use in clinical trials and interventional studies aimed at halting progression of degenerative changes.” 

Limitations of the study include variability of human annotations and a single center study sample and single vendor machine.


De Silva T, Jayakar G, Grisso P, et al. Deep-learning based automatic detection of ellipsoid zone loss in SD-OCT for hydroxychloroquine retinal toxicity screening. Ophthalmol Sci. Published September 24, 2021. doi:10.1016/j.xops.2021.100060